Memrise Replaces Linear Lessons With Three-Layer AI Framework

Memrise以三层AI架构取代线性课程 部署可动态调整上下文与路由的专用智能体，是在向7500万用户规模化推广个性化教育的同时保持真实互动参与度的关键。 Steve Toy介绍了Memrise解决大规模AI个性化问题的方法：摒弃僵化、“千篇一律”的课程体系，转向由用户真实目标驱动的原子化语言学习模型。他提出了以提示词（Prompt）、上下文（Context）和协同（Coordination）为核心的三层架构，阐释了专用多智能体系统如何在语言学习、营养、销售培训和旅游等场景中展开动态交互。通过视频模拟与类比，Toy表明将领域专业知识转化为结构化的多智能体协同，是打造无缝且高度个性化用户体验的关键。 Memrise 首席执行官 Steve Toy 分享了其团队在安永利用 AI 将 400 工时的税务分类任务缩减至 15 分钟的经验；在转型为 AI 驱动平台后，Memrise 于 2024 年被《时代》杂志评为全球第二大教育科技公司。 (0:04) Memrise 摒弃了传统一刀切的线性教学大纲，围绕旅游、家庭沟通及职场发展等真实动机将词汇库“原子化”，并借助数十万条母语者真实视频帮助用户掌握地道发音与俚语。 (0:25) Toy 提出了由 Prompt（提示）、Context（上下文）与 Coordination（协调）组成的三层架构，强调在现代 AI 技术栈中，最大的业务价值在于将领域知识注入应用层以实现大规模个性化。 (0:41) 系统将任务细分给词汇、发音、听力和会话等专用智能体，通过算法编排动态响应用户的实时表现与背景数据，从而避免无语境 AI 带来的荒谬错误。 (16:49)

Steve Toy, CEO of Memrise, argues that delivering effective AI personalization at scale requires businesses to move beyond rigid, one-size-fits-all curricula by decomposing their operations into specialized agents, real-time contextual data streams, and disciplined coordination logic. Atomizing Curricula Around Real-World Goals (0:25) Traditional education systems and legacy digital applications enforce rigid, linear curricula because they are structured to guide an entire cohort toward standardized tests or uniform media consumption. Toy contends this model fails language learners because adult motivations are contextual and practical rather than academic. Real-world learners require targeted language skills for specific outcomes, such as Turkish cab drivers and Peruvian waiters seeking to increase tips from tourists, or individuals learning conversational vocabulary to communicate with foreign-language in-laws. To accommodate diverse learner motives, Memrise atomized its content database into modular word lists, enabling users to acquire vocabulary in any sequence they choose rather than following a fixed syllabus. To ground modularity in authentic communication, the platform pairs these customized lists with hundreds of thousands of native speaker video clips. This multimedia foundation exposes users to authentic accents, natural speech truncations, and local slang that standardized textbooks omit. The Three-Tier Stack and Improv Analogy (0:41) Toy delineates the modern AI landscape into three layers: the foundation layer (LLM providers such as OpenAI and Anthropic managing capital-intensive infrastructure), the middle tool layer (developer toolsets such as LangChain, Cursor, Windsurf, and Weaviate), and the application layer. Toy cautions against starting product development at the middle layer, warning that companies risk acting like a hammer in search of a nail. Instead, sustainable differentiation occurs at the application layer, where domain expertise directly structures user experience. To guide application development, Toy introduces the triad of Prompt, Context, and Coordination, comparing it to an improv comedy performance. In this analogy, a prompt serves as the initial premise (such as opening a coffee shop on Mars rather than a generic coffee shop), context represents dynamic audience feedback that alters the scene in real time, and coordination reflects seamless actor interaction. Toy notes that tight multi-agent coordination produces compelling interactions akin to seasoned performers like Key & Peele, whereas uncoordinated systems resemble chaotic, low-quality improv. Multi-Agent Orchestration and Context Cascades (16:49) At scale, effective personalization requires decomposing complex user interactions into specialized, collaborating agents. Within language learning, Memrise deploys distinct agents for vocabulary curation, pronunciation refinement, listening comprehension, and conversational practice. Context—such as learner goals, cumulative vocabulary count, and recorded mispronunciations—cascades across these agents. If a learner demonstrates pronunciation mastery, the coordination layer automatically bypasses the pronunciation agent and advances the user directly to conversational practice or novel vocabulary. This multi-agent pattern extends to external domains, including nutrition, travel planning, and enterprise sales training. Without shared contextual data, AI systems generate absurd or unsafe outputs, such as recommending squat jumps to an individual in a leg brace or peanut noodle salad to an allergic diner. When properly coordinated, context cascades between specialized agents—such as meal planning, ingredient shopping, and habit tracking—allowing the system to substitute ingredients, alter fitness routines around injuries, or tailor sales objection simulations to an employee's historical call weaknesses. Translating Domain Expertise into Sequencing Logic (19:22) Toy asserts that the ultimate challenge for enterprise leaders is formalizing domain knowledge into algorithmic sequencing. Companies must dissect their business into discrete domain agents, specify the exact contextual inputs each agent requires, identify data origin points, and establish explicit rules governing agent interactions. This sequencing architecture operationalizes human domain judgment into mathematical workflows. According to Toy, mastering this coordination layer separates functional, context-aware platforms from disjointed systems that deliver generic, embarrassing user experiences. Toy emphasizes that enduring competitive advantage in AI does not stem from foundation models or generic tooling, but from how effectively organizations translate proprietary domain expertise into multi-agent coordination. A lingering challenge for enterprise leaders remains the systematic capture and routing of real-time contextual data across disparate agents without creating prohibitive software complexity.

Memrise首席执行官Steve Toy指出，企业在应用层实现大规模个性化AI的核心，在于将行业领域专业知识转化为“提示词（Prompt）、上下文（Context）与协同（Coordination）”三元架构。通过将复杂业务流程拆解为专能智能体并依据实时情境动态编排，AI产品能够摆脱通用模板的线性束缚，兼顾精准度与规模化交付。 商业模式重构与技术栈分层定位 (0:04) Steve Toy将其企业级AI实践追溯至2017年在安永（EY）的经历。当时其团队将一项耗时400工时的税务分类流程缩减为15分钟的自动化计算。面对原有按工时计费模式被打破的恐慌，Toy建议客户转向固定项目收费，价格设定为历史最低年度账单的90%，从而在竞争对手跟进导致利润摊薄前锁定丰厚利润。Toy强调AI并非“魔法”而是“数学”，竞争者的技术追赶必然会迫使初期的高额溢价回归至有利于消费者的水平。 基于此经验，Toy梳理了现代AI的技术栈三层结构：第一层是底层大模型（如OpenAI和Anthropic），涉及巨额资本与高精尖基础设施，普通企业无需涉足；第二层是中间工具层（如LangChain、Weaviate、Cursor等编排与检索工具），Toy警告企业切勿盲目从这一层起步，以免陷入“拿着锤子找钉子”的被动；第三层则是直接面向用户的应用层。Toy认为，只有在应用层将特定行业领域的深厚知识与模型工具有机融合，企业才能真正创造用户感知的价值，2024年Memrise被《时代》周刊评为全球第二大教育科技公司正得益于此层面的深耕。 提示词、上下文与协同的三元机制 (7:00) Toy提出了在应用层落地AI的核心架构方法论——“提示词、上下文与协同”，并用即兴喜剧表演进行了生动类比。提示词相当于演员拿到的基础前提，若设定过于泛化（例如“两个人在排队买咖啡”）则剧情乏味，若设定极具特异性（例如“在火星上开咖啡馆”）便能搭建坚实基础；上下文则是现场观众实时抛出的动态输入（例如“重力装置失灵”），它赋予基础设定独一无二的时空现实；协同则是演员之间的默契互动与接力，协调顺畅时能带来顶级喜剧组合般的体验，缺乏协调则沦为业余闹剧。 在Memrise展示的系统交互模拟中，这一机制体现得尤为明确：当虚拟用户Brenda表示学习西班牙语是为了与姻亲沟通，且提到岳父是个足球迷时，系统不仅将基础对话从死板的语法教程调整为家庭社交用语，更实时注入与足球相关的词汇，随后自主协同调度发音练习工具和相关的母语者真实视频片段。这一过程展现了多智能体系统在捕捉到动态上下文后，如何精准重构交付路径。 语言学习原子化与自适应路由 (0:25) 传统教育和传统移动应用通常强推刚性、线性的学习路径，原因在于教师需要让整个班级应付标准化考试，或应用软件受制于单一的媒体内容库。然而Toy指出，现实中用户的学习动机高度差异化，例如土耳其出租车司机或秘鲁餐厅服务员学习外语是为了增加小费收入，跨国联姻者则是为了融入家庭社交。为此，Memrise将词典库进行“原子化”拆解，允许学习者完全依据自身诉求按任意顺序组合词汇，并依托数十万条真实母语者视频切片，还原带有俚语、吞音和真实语调的原生态语境。 在后台架构上，系统由多个高度垂直的智能体各司其职：词汇智能体负责依据目标挑选词组，发音智能体矫正发音细节，听力/视频智能体在庞大的视频库中检索主要由用户已知词汇构成的片段，对话智能体则提供低压力的模拟实战练习。中央协调逻辑会根据用户实时表现动态调整流转路线，例如当学习者发音表现完美时，系统会自动跳过发音智能体，直接将其引导至新词汇学习或高阶会话练习，实现全流程的自适应调度。 领域专长算法化与跨行业扩展 (16:49) 多智能体协同的架构模式不仅适用于语言学习，还可拓展至旅行规划、营养管理及企业销售培训等众多场景。在旅行场景中，行程智能体、餐饮智能体、活动智能体和预订智能体相互配合，其中餐饮智能体必须精准读取用户的过敏史等深度上下文，避免向贝类过敏者推荐海鲜餐厅；在企业销售培训中，话术教学、异议处理、模拟对练与绩效追踪等智能体，则依托客户画像与销售人员的历史弱点数据进行定制化实训。 Toy通过对比演示强调，缺乏上下文的AI会产生荒谬甚至危险的建议，例如向花生过敏者推荐花生凉面，或向腿部打着石膏的用户建议做深蹲跳跃；而拥有完备上下文的AI则能根据用户身体损伤情况，灵活调整为针对上肢的锻炼方案。因此，企业构建个性化AI的终极挑战在于将行业领域专长转化为清晰的算法逻辑：明确拆解领域内所需的各个专能智能体、定义各环节必需的上下文数据流向，并严谨编排智能体之间的触发时序。 Steve Toy总结认为，实现数千万人规模的AI个性化并非依赖大模型自身的通用能力，而是一项严谨的软件与算法工程。企业的核心壁垒不在于底层算力，而在于能否将自身的垂直领域专长转化为细分智能体、高质量上下文链路与自适应编排逻辑的结合体；若忽视这种深度的业务拆解与协同控制，AI系统不仅无法实现真正的个性化，反而会在交互中暴露出严重的逻辑断层。

## Background 本视频由 Ai4 频道于 2025 年 8 月 24 日发布在 YouTube 平台上。视频记录了 Memrise CEO Steve Toy 的演讲，简介指出其重点探讨了 Memrise 如何通过三层 AI 架构处理 7500 万用户的交互数据，攻克大规模语言学习的个性化难题。 ## TL;DR Steve Toy 分享了他在生产环境中应用 AI 的历程，并重点介绍了 Memrise 如何打破传统固定课程体系，将词汇“原子化”以实现按需学习。演讲提出了以提示词（Prompt）、上下文（Context）和协调（Coordination）为核心的 AI 架构，以即兴表演剧团为喻，阐述了如何通过多智能体协作打造个性化学习体验。此外，他还展示了该架构在语言学习、健康营养及旅游规划等领域的应用对比，强调将领域专业知识转化为数学算法是创造优质 AI 体验的关键。 ## Atmosphere & style 演讲整体氛围专业且幽默，现场感十足，演讲者语言生动、条理清晰。Steve Toy 通过安永税务自动化的亲身经历和“即兴表演剧团”的比喻，将复杂的 AI 理论转化为接地气且易于理解的干货。演示视频中好坏案例的鲜明对比不仅增添了趣味性，也让听众在轻松气氛中深刻体会到精准上下文对 AI 产品体验的决定性影响。 ## Key takeaways - [00:04](https://www.youtube.com/watch?v=S16EsdWJYtc&t=4s) Steve Toy 结合在安永（EY）自动化税务流程的早期经验指出，AI 在商业中的应用并非“魔法”而是“数学”，其超额利润虽会随竞争递减，但能切实提高效率并降低成本。 - [00:25](https://www.youtube.com/watch?v=S16EsdWJYtc&t=25s) 语言学习是一个复杂的“多变量问题”，用户学习语言通常是为了实现具体现实目标（如旅游、职场沟通、增加小费收入等），而非单纯学习语言本身。 - [00:25](https://www.youtube.com/watch?v=S16EsdWJYtc&t=25s) Memrise 打破了传统课程体系和词典架构，将词汇“原子化”，允许用户根据自身目标以任意顺序学习，并结合真实母语者视频以展现真实的口语表达。 - [00:25](https://www.youtube.com/watch?v=S16EsdWJYtc&t=25s) 在 AI 产品构建中，“Prompt（提示词）、Context（上下文）与 Coordination（协调）”是实现大规模个性化体验的核心三大要素。 - [05:14](https://www.youtube.com/watch?v=S16EsdWJYtc&t=314s) AI 技术栈可划分为基础层（大模型与硬件）、工具层（IDE与RAG等工具）和应用层，企业应聚焦应用层，切忌为了使用工具而盲目套用工具。 - [05:14](https://www.youtube.com/watch?v=S16EsdWJYtc&t=314s) 将“Prompt-Context-Coordination”框架类比为“即兴表演剧团”：Prompt 确定基础设定，Context 引入现场变量，Coordination 则决定团队配合与流程调度。 - [11:06](https://www.youtube.com/watch?v=S16EsdWJYtc&t=666s) 在语言学习架构中，拆分出的词汇、发音、听力与会话等 Agent 需要结合用户的学习动机、掌握词汇量及历史偏好等 Context 进行精准调度。 - [11:06](https://www.youtube.com/watch?v=S16EsdWJYtc&t=666s) 精准的 Coordination（协调调度）与智能体分流是领域专业知识（Domain Expertise）的体现，也是实现大规模个性化（Personalization at Scale）的核心关键。 - [11:06](https://www.youtube.com/watch?v=S16EsdWJYtc&t=666s) Agent 架构同样适用于健康营养、旅游规划及 B2B 销售培训等领域，通过将上游 Agent 的输出转化为下一个 Agent 的 Context，可实现自动化协作。 - [16:49](https://www.youtube.com/watch?v=S16EsdWJYtc&t=1009s) 通过向 Agent 注入个性化上下文（如健康禁忌、过敏史或用户过往薄弱点），才能避免推荐错误方案，提供真正精准的智能化体验。 - [18:35](https://www.youtube.com/watch?v=S16EsdWJYtc&t=1115s) 对比演示表明，缺乏 Context 的 AI 容易给出危险或无逻辑的建议（如建议过敏者吃花生），而引入高质量 Context 则能带来贴心且合理的方案。 - [19:22](https://www.youtube.com/watch?v=S16EsdWJYtc&t=1162s) 构建个性化 AI 系统需要完成三步：拆解业务领域以明确 Agent 职责、梳理并注入高质量 Context 数据源，以及设计合理的执行算法逻辑。 ## Detailed walkthrough 以下是根据视频 Part 1 内容整理的详细逐段记录与分析： ### [00:04](https://www.youtube.com/watch?v=S16EsdWJYtc&t=4s) 从税务自动化到生产级 AI：Steve Toy 的 AI 历程与 Memrise 的定位 Memrise 的 CEO Steve Toy 讲述了他运用 AI 的起点：2017 年在安永（EY）工作期间，他刚将上一家公司卖给安永。入职约九个月时，团队成功将原本需要 400 个工时的税务分类工作，缩短为仅需 15 分钟的计算机运行程序。然而，当他们向业务负责人演示这一成果时，对方却“吓坏了”——不是因为程序不起作用，而是因为太有效了！安永当时是按“工时与材料（Time and Materials）”计费的。Steve 和团队向他解释并建议转变商业模式：改为项目制计费，每月向客户收取固定费用（设定为过去 12 个月最低账单的 90%），从而在执行过程中直接套现数十万美元的毛利与净利润。当这位业务负责人算清这笔账后，惊叹道：“这简直是魔法！”Steve 回应道：“**不，这不是魔法，这是数学。**” 既然是数学，就必然会被他人效仿。如今，当年在安永合作该项目负责 AI 的同事已开办了一家 AI 会计事务所，原有的超额利润虽因市场竞争而消逝，但客户却得以用更少的钱完成同样的工作。 这段经历开启了 Steve 在实际生产环境中应用 AI 的旅程。此后，他管理过健身、营养、工具、VPN 以及休闲游戏等多类 App，如今则执掌语言学习平台 Memrise。他指出，语言学习是一个极其复杂的“多变量问题（Multivariate problem）”，而 Memrise 的目标是通过 AI 加速这一过程，提供个性化体验以帮助每个人取得成功。2024 年，《时代》杂志（Time magazine）将 Memrise 评为全球第二大教育科技公司（EdTech Company），正是归功于其后台部署的 AI 技术。Steve 强调，Memrise 现已不再只是一款简单的 App，而是一个完整的语言学习平台。 ### [00:25](https://www.youtube.com/watch?v=S16EsdWJYtc&t=25s) 破解个性化难题：以目标为导向的原子化词库与真实母语者内容 介绍 Memrise 的底层机制时，Steve 表示用户学习语言的终极目的从来不是“为了学语言本身”，而是为了实现更具体的现实目标： - 获得更好的旅游体验； - 与刚结婚的跨国家庭中的岳母/婆婆沟通； - 与新公司的同事流畅交流； - 土耳其的出租车司机希望能和乘客沟通，从而赚更多钱； - 秘鲁的餐厅服务生希望能更好地服务外国顾客，从而获得更多小费。 传统教育和大多数 App 总是强制所有人按照统一的课程体系（Curriculum）学习，其本质是“把所有学生推向同一次考试或同一段视频内容”。Memrise 则打碎了传统词典，将词汇“原子化（Atomized）”，允许用户根据自身目标，以任意顺序学习所需的任何词汇。系统会将用户选择的不同词汇组合输入到 AI Agent（智能体）中，由 Agent 协作协助用户达成口语目标。同时，Memrise 完整保留了其核心特色——包含数十万段由真实母语者（Native speakers）录制的视频内容，让用户体验包含连读、省略及俚语的真实口语表达。 为了帮助与会者将这套经验应用到各自的业务中，Steve 提出了本次演讲的核心框架：**Prompt（提示词）、Context（上下文）与 Coordination（协调）**。此外，他还预告了下午 4:30 在 Marquee 大宴会厅举行的 Ask Me Anything (AMA) 问答环节，邀请现场听众前来挑战。 ### [05:14](https://www.youtube.com/watch?v=S16EsdWJYtc&t=314s) 聚焦应用层：AI 技术栈与“即兴喜剧”框架类比 Steve 简要拆解了 AI 的技术栈（Tech Stack），以明确本次讨论的重点范畴： ``` +-----------------------------------------------------------------+ | Foundation Layer | | (OpenAI, Anthropic, LLMs, Hardware & Power) | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | Tools Layer | | (Cursor, Windsurf, LangChain, Weaviate / RAG, etc.) | +-----------------------------------------------------------------+ | v +-----------------------------------------------------------------+ | App Layer | | (User Interface, Domain Expertise, Personalization at Scale) | +-----------------------------------------------------------------+ ``` - **基础层（Foundation Layer）**：由 OpenAI、Anthropic 等巨头垄断。该领域涉及数亿美元薪酬的争夺、国会听证会答辩以及核反应堆供电等议题。普通企业无需关心其底层运作，只需了解模型能为业务提供什么帮助。 - **工具层（Tools Layer）**：包括 Cursor、Windsurf 等 IDE，LangChain 等 Agent 编排工具，以及 Weaviate 等 RAG 工具。Steve 警告不要“为了用工具而学工具”，否则会陷入“拿着锤子找钉子”的误区。 - **应用层（App Layer）**：用户真正停留、感受“魔法”的地方。企业需要在此结合自身领域知识（Domain Expertise），利用技术框架为用户创造价值。 为了让“Prompt、Context、Coordination”框架更直观，Steve 用**即兴表演剧团（Improv Troop）**做比喻： 1. **Prompt（提示词 / 设定）**：类似于剧团演员与基本前提。若前提过于宽泛（如“一群人去买咖啡”），过程就会极其无聊——就像在拉斯维加斯美食广场的星巴克排队半小时买一杯昂贵的拿铁，刷着手机怀疑人生；但如果设定为“在火星上开一家咖啡馆”，表演立马变得有吸引力。 2. **Context（上下文 / 现场变量）**：类似于观众给出的突发设定（例如“火星咖啡馆的人造重力发生器突然坏了”或“老板要求在做咖啡时引入 AI”），这能让通用的 Prompt 变得独一无二。 3. **Coordination（协调 / 团队配合）**：演员之间的默契配合。优秀的协调（如 Key & Peele 或 Abbott and Costello 的经典《谁在第一垒》）能创造奇迹；而缺乏协调的表演就像偏远小镇条形码商场里的蹩脚演出。 随后，现场播放了团队成员 Emma 仅用几个小时制作的模拟演示视频：展示了一名名为 Brenda 的用户为了“见岳父/公公”而学习西班牙语，系统自动为其生成家庭与足球相关词库、纠正发音，并推送母语者视频进行战前特训的全过程。 ### [11:06](https://www.youtube.com/watch?v=S16EsdWJYtc&t=666s) 实践案例：语言学习与健康营养领域的 Agent 架构 在演讲后半段，Steve 详细拆解了如何将“Prompt - Context - Coordination”框架应用到具体的业务场景中： #### 1. 语言学习场景（Language Learning） - **Vocabulary Agent（词汇 Agent）**： - *Prompt*：挑选达成目标所需的词汇。 - *Context*：用户的学习动机（如前往巴塞罗那旅游且喜爱足球与美食 vs. 与喜爱针织的岳母沟通）。 - **Pronunciation Agent（发音 Agent）**： - *Prompt*：纠正发音，使其接近母语者。 - *Context*：获取用户在其他环节中发音不佳的历史数据，进行专项强化。 - **Listening & Video Agent（听力与视频 Agent）**： - *Prompt*：从数十万视频库中筛选适合当前水平的内容。 - *Context*：结合用户已掌握的词汇量（即便两人都掌握 250 个词，只要具体词汇不同，推送的视频就不同）及偏好标签（如频繁选择体育视频）。 - **Conversation Agent（会话 Agent）**： - *Prompt*：模拟真实对话场景（如订咖啡或餐厅预订）。 - *Context*：用户的学习目的（Why）以及词汇量储备，以此动态调节对话的复杂程度。 - **Coordination（协调）**：若用户发音完美，系统便无需将其引导至发音 Agent，而是直接推进至新词汇和对话练习。**这种精准分流正是领域知识的体现，也是实现大规模个性化（Personalization at Scale）的核心。** #### 2. 营养健康场景（Nutrition） - **Meal Plan Agent**：根据目标（减重、降胆固醇、增重、预防心脏病等 Context）制定一周饮食计划。 - **Shopping Agent**：接收 Meal Plan Agent 输出的计划作为 Context，自动生成针对性的采购清单。 - **Food Swap Agent**：了解用户的健康目标及日常饮食习惯（Context），在发现用户食用切多斯芝士条（Cheetos）时，主动推荐海带芯片（Kelp chips）等健康替代品。 - **Progress Agent**：综合输入（饮食记录、睡眠数据）与输出（体重变化、胆固醇指标等 Context），评估健康改善效果。 #### 3. 旅游规划场景（Travel） - 部署专门掌握特定区域知识的 **Location Agent**（甚至细分为巴塞罗那 Agent、马德里 Agent 或伊比萨 Agent）； - 结合专门寻找当地美食的 **Food Agent**，以及挖掘特色体验与本地人活动的 **Activity Agent**，共同构建个性化行程。 ### [016:49] 智能体架构中的个性化上下文与应用场景拓展 Steve Toy 进一步阐述了如何通过个性化上下文来赋能不同的专用智能体（Agents）。以旅游规划为例，系统中可以包含旅游智能体、美食智能体、活动智能体以及预订智能体等。每个智能体都在等待特定的指令，但要提供真正优质的服务，就必须注入个性化上下文。美食智能体必须清楚用户的饮食喜好与禁忌，绝不能将海鲜过敏的顾客推荐去龙虾小屋（lobster shack）；活动智能体需要掌握用户偏好博物馆还是体育运动；而预订智能体则在用户确定意向后直接完成预订，免去繁琐流程。 这种基于“待办任务”（Jobs-to-be-done）的智能体架构不仅适用于消费级 App，在 B2B 场景（如销售培训）中同样大有可为。销售培训系统可以拆解出销售技巧、异议处理、模拟演练和绩效评估等智能体。通过输入潜在客户的需求偏好、常见异议点以及销售人员过往的薄弱环节等上下文，性能智能体甚至可以直接调取并分析通话录音，从而提供量身定制的培训反馈。 ``` +-----------------------------------------------------------+ | 个性化上下文 (Context) | | (用户喜好/禁忌/历史表现/潜在客户需求/通话录音等) | +-----------------------------+-----------------------------+ | v +-----------------------------------------------------------+ | 专用智能体 (Agents) | | [美食智能体] [活动智能体] [销售技巧智能体] [预订智能体] | +-----------------------------+-----------------------------+ | v +-----------------------------------------------------------+ | 精准推荐与自动化执行 | +-----------------------------------------------------------+ ``` - 美食智能体必须掌握用户的过敏史与偏好，避免做出不合逻辑的推荐。 - 销售培训领域可结合历史通话录音，精准识别销售人员的不足之处。 ### [018:35] 优质上下文与劣质上下文的实际对比演示 为了直观地展示提示词（Prompts）中引入优质上下文与劣质上下文的巨大差异，演讲现场播放了来自 Emma 的对比演示视频。在缺乏上下文的糟糕示例中，AI 面对用户的求助，竟向对花生过敏的用户推荐做“花生面条沙拉”（peanut noodle salad），向戴着腿部支具的用户建议做“45秒跳蹲”（squat jumps），引发了极度尴尬与不适。 而在注入精准上下文的优质示例中，AI 智能体表现得极为贴心与智慧： > “Hey Mia, you're at 10% protein for the day and you've been working out a bunch. I know popcorn for dinner again sounds good, but try salmon and greens.” AI 能够根据 Mia 当天仅摄入 10% 蛋白质且运动量较大的数据，精准劝阻她晚上再吃爆米花，转而推荐三文鱼和蔬菜；针对戴着支具的 Dan，AI 也能识别出他昨天刚练完胸部，适时调整训练计划为手臂练习，既安全又合理。 - 缺少上下文的 AI 容易给出危险或毫无逻辑的建议（如建议过敏者吃花生、建议受伤者做跳蹲）。 - 具备高质量上下文的 AI 能结合用户当天的营养摄入、体能消耗及身体状况提供个性化方案。 ### [019:22] 业务落地挑战与总结展望 演讲最后，Steve Toy 提出了针对所有企业和业务负责人的落地挑战。他强调，要构建出真正个性化的 AI 系统，必须完成三个关键步骤：第一，拆解业务领域要素，明确需要哪些智能体（Agents）；第二，梳理各个环节所需的上下文信息，并找到数据的获取来源；第三，设计合理的执行顺序（Sequencing），即算法逻辑。 Steve Toy 形象地指出，这种将业务领域的专业知识转化为数学算法的过程，正是 AI 产生“神奇魔力”的源泉。做好了，你的产品就是精彩绝伦的《基和皮尔》（Key & Peele）喜剧段子；做不好，就像是商业街里蹩脚的即兴表演队。最后，他邀请有兴趣的听众在下午 4:30 前往 Marquee Ballroom 继续交流，或扫描现场二维码与他和 Emma 联系，希望大家能将这些可落地的实操经验应用到自身业务中。 - 构建个性化 AI 的三要素：业务领域拆解（智能体）、上下文数据注入（个性化）、执行顺序设计（算法）。 - 将领域专业知识（Domain Expertise）转化为数学算法，是决定 AI 产品成败的关键。 ## Notable quotes > [00:04](https://www.youtube.com/watch?v=S16EsdWJYtc&t=4s) 不，这不是魔法，这是数学。 — Steve Toy > [00:25](https://www.youtube.com/watch?v=S16EsdWJYtc&t=25s) 把所有学生推向同一次考试或同一段视频内容 — Steve Toy > [18:35](https://www.youtube.com/watch?v=S16EsdWJYtc&t=1115s) Hey Mia, you're at 10% protein for the day and you've been working out a bunch. I know popcorn for dinner again sounds good, but try salmon and greens. — AI Assistant ## Mentioned Steve Toy, Memrise, 安永 (EY), OpenAI, Anthropic, Cursor, Windsurf, LangChain, Weaviate, Emma, Key & Peele, Abbott and Costello ## Video description Steve Toy, CEO, Memrise Abstracts Beyond One-Size-Fits-All: The AI Architecture That Actually Personalizes Learning" Most AI demos look impressive until you try to deploy them to millions of real users with different languages, learning styles, and goals. At Memrise, we've spent years solving the most complex problem in AI-powered education: true personalization at scale. I'll reveal the three-layer AI architecture that transforms 75 million user interactions across 200+ language pairs into perfectly tailored learning experiences. You'll discover why most personalization attempts fail, the data patterns that actually predict learning success, and the counterintuitive approaches we use to make every lesson feel like it was designed specifically for each learner. This isn't theory—it's battle-tested insights from the world's most personalized language learning platform. [Source](https://www.youtube.com/watch?v=S16EsdWJYtc)

[0:04] Steve Toy: 不，这不是魔法，这是数学。 [0:25] Steve Toy: 把所有学生推向同一次考试或同一段视频内容 [18:35] AI Assistant: Hey Mia, you're at 10% protein for the day and you've been working out a bunch. I know popcorn for dinner again sounds good, but try salmon and greens.

[0:04] Steve Toy: No, it's math [4:01] Steve Toy: a proverbial hammer looking for a nail [7:00] Learner: Hola, can you help me improve my Spanish? [7:00] Learner: How do you say 'the cat ate an apple'? [7:00] Learner: This isn't going to help me meet my in-laws. [7:00] Agent: So, Brenda, why are you learning Spanish? [7:00] Brenda: To meet my in-laws. [7:00] Agent: Got it. I'll add family, father-in-law, and some 'just-in-case' compliments to your word list. [7:00] Agent: for when you pretend to know what just happened [17:58] User: I'm allergic to peanuts [19:22] Steve Toy: This is what when you do it well, you're Key & Peele. And when you don't, you're a bad improv troupe in the strip mall.

[0:04] All right. Hello everybody. [0:06] So, let me just get kicked off here [0:08] talking about where I came from here. [0:12] And we'll start with 2017 when I was at [0:15] EY. I just sold the previous company I [0:18] was running to EY. And about nine months [0:20] in, we managed to put 400 man hours of [0:25] tax classification work into a 15minute [0:28] computer run. And when we shared that [0:31] with the business owner, he freaked out [0:33] a bit. Not because it didn't work, but [0:35] because it does. And he builds on a time [0:37] builds on a time and materials basis. So [0:40] we explained to him the thing you need [0:42] to do here is go project basis. Go to [0:45] your clients and say that you're going [0:46] to charge them the same flat fee every [0:48] month. 10% less than the lowest invoice [0:51] that you had in the last 12 months and [0:54] then bank hundreds of thousands of [0:56] dollars of gross profit and net profit [0:58] along the way. And once we explained how [1:01] that could work, he literally sat there [1:04] seeing all of that profit and said, [1:05] "Man, this stuff is magic." And the [1:08] answer to him was, "No, it's math." And [1:10] because it's math, it's going to get [1:13] followed up by other people. Fast [1:16] forward to today and the uh AI the guy [1:19] that ran AI that I worked with to make [1:21] that happen at EY has opened up an AI [1:24] accounting firm. That margin is gone but [1:27] to the benefit of customers who are [1:29] paying less money for doing the same job [1:31] and that's where my journey on using AI [1:35] in production started. Since then, I've [1:38] gone and I've run a whole lot of apps, [1:41] fitness apps, nutrition apps, utility [1:43] apps, VPN apps, casual gaming apps, and [1:46] now I'm running uh Memorize, which is a [1:49] language learning app where we help [1:51] people learn to speak a language to [1:53] further their goals. And what we're [1:56] trying to do is take the very intense [1:58] multivariate problem that is language [2:00] learning and boost it with AI so that we [2:04] can give a personalized experience and [2:05] help everybody succeed. And so far we're [2:08] doing okay with it. In 2024, Time [2:11] magazine said we were the number two [2:13] edtech company in the world based on the [2:15] backs of all of this AI that we've been [2:17] implementing for this purpose. And now [2:19] we can say that we're for sure no longer [2:22] an app, but a language learning [2:23] platform. [2:25] So to give you a little bit of what we [2:27] do, which is important for the framework [2:28] that I'm going to give you here today, [2:30] I'll tell you that memorize tries to [2:33] help people learn a language to [2:35] accomplish their goals. And their goal [2:37] is not to learn a language. Their goal [2:40] is to travel better or to talk to their [2:43] mother-in-law in the family that they [2:44] just married into or talk to their [2:47] co-workers in the company that they just [2:49] joined. Or you're a cabby in Turkey and [2:51] you make more money when you can speak [2:53] to the people who come to get in your [2:55] cab. Or you're a waiter in Peru and you [2:57] can make more money when you can speak [2:59] to the people that you're waiting on. So [3:01] those are the goals that people have and [3:03] those goals are defined by the words [3:06] that they learn. And so we start our [3:09] journey with uh spec specific word [3:12] lists. We allow people to learn whatever [3:14] words they want to learn in whatever [3:15] order. That has never been the way that [3:18] anybody has done that. You've been in [3:20] school and you learn curriculum. You go [3:21] to apps, you learn curriculum and you [3:23] learn it not because that's the best way [3:25] to learn it that way, but because [3:27] teachers are marching all the kids to [3:29] the same test. Our apps are marching [3:31] everybody to the same piece of video [3:33] content, but that's not the best way to [3:35] learn. So, we've atomized our [3:36] dictionaries and help people learn their [3:37] goals. And then we take every different [3:40] set of words and run them into AI agents [3:43] that help them do their job, help you [3:46] accomplish the goal of being able to [3:47] speak a language. And it's different for [3:50] everybody what we do with that mix. And [3:53] finally, we're doing it all while [3:55] keeping the native speaker content that [3:58] is defined memorized from the beginning. [4:00] We have hundreds of thousands of native [4:03] speaker content in our app so that you [4:05] can hear the language as it's spoken by [4:07] natives with all the truncation and [4:09] slang and everything involved. So that's [4:12] what we do and that's why we I have the [4:15] ability to give you the framework that [4:18] we're going to give you today here. So [4:19] my goal right now is to give you a [4:21] framework that you can apply to your [4:23] businesses. I really sincerely want to [4:25] make this the most useful presentation [4:27] that you have this entire three days and [4:30] I'm going to give you a challenge at the [4:32] end of it and we'll talk about that. But [4:33] right now, a framework that can apply to [4:35] your business. In order to do that, I [4:38] need to focus in on the right area. So, [4:40] I'll give you a quick area of what part [4:42] of the technology we're focusing on and [4:44] then I'm going to give you the [4:45] framework, which is as simple as this. [4:46] Prompt, context, coordination. That's [4:49] it. I'm going to say it over and over a [4:51] million times. I'm going to give you a [4:52] zillion examples, but that is the [4:54] framework. And then, as I said, I'm [4:56] going to give you a challenge at the [4:57] end. And um that challenge can be [4:59] followed up upon because I'll be [5:02] speaking at 4:30 in the marquee ballroom [5:05] on an ask me anything. So, you can bring [5:07] all your questions to that session and [5:10] test me live on stage. So, um the AI [5:14] tech stack in order to focus us in. This [5:16] is basically the AI tech stack. If you [5:19] were at the keynote this morning, [5:20] Catalina talked about the foundation [5:22] layer is the is the first area in her [5:24] speech. She talked about the tools area [5:26] as well, but this is a tech stack. The [5:29] foundation layer, that's all the big [5:32] dogs, open AI, anthropic, and so on. If [5:34] you are currently on or are negotiating [5:37] a hundred million dollar salary, well, [5:39] you're you're in that layer. Or if [5:41] you're appearing in front of Congress [5:43] trying to explain why you're not [5:44] destroying the world, you're in that [5:46] area. or if you're casually thinking [5:48] about how to spin up a nuclear reactor [5:50] for power, you're in this area. Most of [5:53] us don't need to know how this works, [5:55] but we do need to know how the models [5:58] can help our business. So, we have to [6:00] understand our business and what we need [6:02] from it um first, but we don't need to [6:04] spend a lot of time there. The middle [6:06] layer is all the tools. This is the [6:09] idees like cursor and windsurf. These [6:12] are uh agent orchestration things like [6:14] lang chain. These are rag tools like [6:16] weev8. These are the tools. But you [6:18] don't want to go and learn a particular [6:20] tool first. Otherwise, you're going to [6:22] be the proverbial hammer looking for a [6:24] nail. You got to figure out your [6:25] business and figure out what tool will [6:28] help you do your thing better. So, we're [6:31] talking about the app layer. The app [6:34] layer is where the users live. The app [6:36] layer is where the users feel the magic. [6:38] The app layer is where you bring your [6:40] domain expertise and all the technology [6:43] and tools that we just talked about to [6:45] the users and you do it with the [6:47] framework that we're going to talk about [6:49] here. Now, this framework I am trying to [6:53] I was trying to figure out all kinds of [6:55] different ways to make this framework [6:56] understandable and I've got a couple and [6:58] I'm going to start with just an analogy [7:00] and the analogy is going to be an improv [7:02] troop. And an improv troop is very much [7:05] like an agent with a basic prompt. You [7:08] got some actors and you got a premise, [7:10] right? So, if you have a bunch of actors [7:12] and the premise is very broad and [7:14] unspecific, it's boring. This is like, [7:16] okay, bunch of people get coffee. Well, [7:19] then you're going to have an experience [7:20] like maybe you guys, like I had this [7:22] morning, maybe you had, right? You're [7:23] standing in a Starbucks at a food court [7:25] in Vegas, sitting there doom scrolling [7:27] on your phone for a half an hour to get [7:29] a latte that is more expensive than a [7:31] top shelf margarita back home, and [7:33] you're just wondering whether you need [7:34] that caffeine that bad. But if you have [7:37] a more specific prompt, right, if you [7:39] have a prompt that um has um you know [7:42] more to it, such as you're opening up a [7:44] coffee shop on Mars, right, then you've [7:47] got something to work with. So in the [7:49] improv area, you got people and the [7:51] premise, which is like an agent and a [7:53] prompt. And then you get context, right? [7:56] And context in the improv area comes [7:58] from the audience. the audience says, [8:00] "Okay, you your gravity machine just [8:03] broke in your coffee shop on Mars, or [8:05] your boss just told you you needed to [8:07] implement AI in the coffee making [8:08] process." So, that would be something [8:11] that makes a common premise or prompt [8:15] unique for that night. So, you can have [8:18] something unique for every audience. And [8:21] then finally, there's the coordination [8:23] amongst all the actors, right? So, when [8:25] that's done well, it's magic. Think, you [8:27] know, Tina Fay and Amy Polar working off [8:29] of each others. Think Key and Peele [8:31] coming up with a crazy premise and just [8:33] running with it. Think Abbott and [8:35] Costello and who's on first. This is [8:37] when it's done well as compared to think [8:40] about improv comedy in a strip mall in [8:43] some, you know, faroff city. Not done so [8:46] well. So, there's different experiences. [8:48] Another way that I can explain what we [8:52] do here and give you um an idea about it [8:55] is to use AI. And I shouldn't say that [8:58] I'm using AI. Um the wonderful and [9:00] talented Emma sitting right over here [9:02] just made a movie for us to make this [9:05] point about what Memorize does, a [9:06] simulation about what Memorize does. And [9:08] I bet you if a guy from Arcana was the [9:11] keynote speaker from Arcana was here, [9:13] he'd be like, "Yeah, that's what I'm [9:14] talking about. She did this in a few [9:16] hours here. [9:23] Ola, can you help me improve my Spanish? [9:26] >> Sure. [9:26] >> How do you say the cat ate an apple? [9:29] >> This isn't going to help me meet my [9:31] in-laws. [9:38] So, Brenda, why are you learning [9:39] Spanish? To meet my in-laws. Got it. [9:42] I'll add family, father-in-law, and some [9:44] just in case compliments to your word [9:46] list. [9:47] What's the word for father-in-law? [9:48] [music] [9:50] >> Nice. You're nailing this. Do you want [9:52] to focus on anything specific today? I [9:54] found out my father-in-law loves [music] [9:56] soccer. Can we learn some sports [9:57] phrases? [9:58] >> Of course. I'll add football, goal, and [10:00] other phrases for when you pretend to [10:02] know what [music] just happened. [10:04] Time for pronunciation. Can you say what [10:06] to play? [10:08] >> Not exactly, but you're getting close. [10:11] Hey, Brenda. Trip soon, and we know [10:13] pronunciation's been tough. These native [10:15] speaker videos might help. Tomorrow's [10:18] the big day. Want to try that phrase [10:20] once more? [10:22] >> Perfect. Your in-laws are going to love [10:24] you. [10:36] there. There you go. Emma deserves your [10:39] praise over there for that one. I'd like [10:41] to say I was talented enough to make it, [10:43] but that was her. So, what we're going [10:45] to do now is we're just going to go [10:46] through prompting context and [10:48] coordination a bunch of times with [10:49] different areas to really land the [10:51] point. And hopefully in some cases I'll [10:53] get close enough to your businesses that [10:55] it will start to really feel [10:57] comfortable. But at least with these [11:00] number of reps, you'll you'll certainly [11:02] know how to take this into your [11:04] businesses. So, we're going to start [11:06] with the language learning space, right? [11:08] And this is not all the agents that we [11:10] have for sure, but it's enough to make [11:12] the point. So, we have a vocabulary [11:13] agent. It picks the words that are [11:15] necessary for you to accomplish your [11:17] goals. So, if you say that, I'm [11:18] traveling to Barcelona and I love [11:20] football and um and food will give you [11:24] different words than if you say, "I'm [11:26] going to talk to my mother-in-law um and [11:29] she loves knitting." So, um a vocabulary [11:32] agent. We have a pronunciation agent. It [11:34] just helps you pronounce words and get [11:36] you closer to native speaking sound. We [11:38] have a listening and video agent that [11:40] goes and looks through our hundreds of [11:42] thousands of videos and figures out [11:44] which videos are made up mostly of words [11:47] that you know so that any three people [11:49] who all have learned 250 words but [11:52] different 250 words can have those [11:54] videos filtered for them and have enough [11:57] of them so that they can go to what they [11:59] like. And then finally, a conversation [12:02] agent. Something that can simulate a [12:04] conversation so you can get some batting [12:06] practice in rats before you go and order [12:08] a coffee in Barcelona. Now, those are [12:11] the prompts. We get them good at doing [12:13] these things. And then we bring them [12:15] context. The context in the vocabulary [12:18] agent is why you learn the language, [12:20] right? And and that can tell us what [12:22] vocabulary we have. Pronunci [12:25] pronunciation agent that uses what [12:28] things have you pronounced poorly in [12:30] interacting with the other things in the [12:32] system and we will use those to bring to [12:35] the pronunciation agent and say these [12:37] are the things you need to practice [12:38] right listening and video agent what we [12:41] need to know is the words that you've [12:42] learned so that we can filter the videos [12:44] that you have a chance of understanding [12:47] and over time if you're constantly [12:48] selecting sports videos then we also can [12:51] filter on that as well if we get that [12:53] signal [12:54] And then finally, the conversation [12:56] agent, like what's your why? If you're [12:57] learning to travel, let's get you [12:59] practicing ordering a coffee or making [13:01] restaurant reservations or talking about [13:03] museums. Um, your why matters. And then [13:06] amount of words that you have allows us [13:08] to dial up or dial down the [13:09] sophistication of the conversation that [13:11] you're having. [13:13] The final part is the coordination part, [13:15] which is how do we pass people around, [13:17] right? If somebody is pronouncing [13:19] everything perfectly, we don't need to [13:20] pass you to the pronunciation agent, [13:22] right? We just keep you know having you [13:23] learn more words and watching more [13:24] videos and have more conversations. So [13:27] this is where everything becomes unique. [13:29] This is where your domain expertise [13:31] lives. This is is what gives you the [13:34] magic. This is how you figure out how [13:37] you make your users satisfied with the [13:41] product that you're giving them. This is [13:42] how you personalize at scale. [13:46] Now I'm going to give you a bunch more [13:48] just like this. All right. Um, I'm going [13:51] to start with nutrition. Um, and we'll [13:53] go from there. I'm not going to do the [13:55] coordination thing over and over again [13:56] because it's the same point. You got to [13:58] coordinate these agents. But in [14:00] nutrition, you might have a meal plan [14:03] agent, right? An agent that knows how to [14:06] make a meal plan and sets it out for the [14:08] week, right? And then you might have a [14:12] shopping agent. And that shopping agent [14:15] is going to make a shopping list. You'll [14:17] have a food swap agent that can give you [14:20] suggestions about what to swap out in [14:22] order to get a different food of a [14:24] similar type in order to help you [14:27] accomplish your objectives and perhaps a [14:29] progress agent that looks at inputs and [14:32] outputs. [14:34] Now, we bring context to it, right? The [14:37] meal plan agent's got to know what your [14:39] goal is, right? Are you trying to lose [14:40] weight? You're trying to lower your [14:42] cholesterol? You're trying to gain [14:43] weight? You're trying to avoid cardiac [14:46] disease. What is your goal? Because the [14:49] meal plan depends on that goal. But once [14:53] you have a meal plan, right, the [14:55] shopping agent now has context. All [14:57] right, there's a plan. I can now put [14:59] together the shopping plan. And so the [15:02] context doesn't always come directly [15:04] from the user. It may come from other [15:08] just like in the pronunciation area, [15:11] right? The food swap agent needs to know [15:13] a your goal, but what foods you usually [15:16] eat. And maybe that comes from just [15:19] using the app for a little while. Maybe [15:21] you tell it, but somehow it knows what [15:23] foods you're eating and can start to [15:25] suggest, you know, given your goals, [15:28] those Cheetos are really not helping you [15:30] further your goal. How about you try [15:33] this kelp chips or something? And and [15:36] then finally, there's the progress [15:37] agent. A progress agent can take in [15:39] information like weight or your [15:41] cholesterol numbers. It can take in [15:43] information about what you've been [15:44] eating. It can take in how you've been [15:46] sleeping. It can take in all kinds of [15:48] things. It depends on what you want to [15:50] feed it in order for it to examine the [15:52] inputs and the outputs and give a [15:54] result. So, bunch of agents prompted to [15:58] do a job given context from the user or [16:01] other agents in the app. [16:04] Travel. Travel is another area. So, in [16:08] the travel area, you might have a [16:09] location agent that is expert in knowing [16:11] all that is going on in a bunch of [16:14] areas. Heck, you might even have a [16:16] Barcelona agent and a Madrid agent and [16:18] an Aiza agent, right? Um, or you might [16:21] just have a Spain a agent. Don't know. [16:23] It depends on your business, but [16:25] something that knows the location. [16:27] Might have a food agent, right? A lot of [16:29] people go traveling and, you know, food [16:32] is a big part of the discovery. So, what [16:34] are the best places? What are the best [16:35] things to eat? Right? You could have an [16:38] activity agent. What are the things that [16:40] are fun to do in this particular area? [16:42] What do the locals do? What what um do [16:45] you enjoy doing and how can you do the [16:46] thing you enjoy in this location? [16:49] Right? And then the reservation agent is [16:51] something that could potentially book [16:53] things if it knows what you want booked. [16:57] So, these are agents that are prompted [16:59] and just sitting waiting to do a job, [17:01] but you still want to personalize them. [17:04] So, when it comes time to personalizing [17:06] them, right, the travel agent needs to [17:08] know what locations have the things that [17:10] you like to do, right? The food agent [17:13] needs to know what food do you like, [17:15] what food do you dislike, what food do [17:17] you not want. They they shouldn't tell [17:19] send you to the lobster shack if you [17:22] have a shellfish allergy, right? That [17:23] just doesn't make sense, right? The [17:26] activity agent also needs to know what [17:28] you like, museums, sports, what have [17:30] you, so that you can give a better [17:32] recommendation. And then once you have [17:35] decided the things that you might want [17:37] to do or the things that you could do, [17:38] well, then you have a reservation agent [17:40] who might be able to just book this for [17:42] you right away so you don't have to go [17:43] through all of the challenges. [17:46] Now these are all sort of apps that [17:49] somehow can you know you can imagine you [17:52] you probably have apps on your phone [17:54] like them but it doesn't have to be just [17:56] in the app world. Let's take something [17:58] like sales training. So in the sales [18:01] training world there's certain jobs to [18:03] be done. In those jobs to be done right [18:06] there is you know sales techniques. Oh I [18:09] see I'm getting the two-minute sign so [18:10] I'm going to go fast here. Sales [18:11] techniques there's objections. There's [18:14] simulations of sales calls and there's, [18:17] you know, performance. So that can be [18:20] fed with the context of, you know, what [18:22] is the prospect, you know, desire? What [18:25] objections might they have? What things [18:28] have has the salesperson not done well [18:30] in the past? And the performance agent [18:32] can take recordings of calls. [18:35] So to make this uh point about what good [18:37] context and bad contacts with prompts, [18:40] let's do one more of Emma's videos here. [18:43] >> [music] [18:45] >> Hey, can you help me improve? [18:46] >> Can you help me improve? [18:47] >> Sure, I'll teach you how to make a [18:48] peanut noodle salad. [music] [18:50] >> I'm allergic to peanuts. Let's crush 45 [18:54] seconds of [music] squat jumps. [18:57] >> I'm wearing a leg brace. Why would you [18:58] suggest that? [19:03] >> Hey, Mia, you're at 10% protein for the [19:05] day and you've been working out a bunch. [19:07] I know popcorn for dinner again sounds [19:08] good, but try salmon and greens. Hey [19:11] Dan, you crushed chest yesterday and the [19:13] brace is still on. Let's keep it smart. [19:16] Arms today. You'll get to flex after. [19:22] So, you get the point here. So, [19:24] hopefully I've given you enough reps on [19:26] this one to give you the challenge, [19:28] right? So, the challenge is look at your [19:30] business and figure out what are the [19:32] elements of your domain. These are your [19:34] agents, right? What are the pieces? And [19:36] you got to break them down into the [19:37] pieces. What context do those pieces [19:40] need and where do you get that [19:41] information? Right? This is the context [19:43] that you can feed the elements of the [19:44] domain. This is where the [19:45] personalization starts. And then of [19:47] course, what sequencing needs to happen [19:50] in order for this all to work. This is [19:52] your algorithm. This is your domain [19:54] expertise turned into math. This is what [19:56] makes the magic. This is what when you [19:58] do it well, you're key and peel. And [20:00] when you don't, you're a bad improv [20:02] troop in the strip mall. So take that [20:05] challenge, apply it to your business. [20:07] Come talk to us at 4:30 in the marquee [20:10] ballroom if you have any questions or [20:12] take the QR codes for both me and Emma [20:15] and we'll stick around afterwards and [20:17] talk to you. But hopefully this gives [20:19] you something that you can action today [20:22] and makes this uh a useful use of your [20:24] time. So, thank you very much. [20:26] appreciate it and uh enjoy the rest of [20:28] the